net/smc: Avoid warning of possible recursive locking

Possible recursive locking is detected by lockdep when SMC
falls back to TCP. The corresponding warnings are as follows:

 ============================================
 WARNING: possible recursive locking detected
 5.16.0-rc1+ #18 Tainted: G            E
 --------------------------------------------
 wrk/1391 is trying to acquire lock:
 ffff975246c8e7d8 (&ei->socket.wq.wait){..-.}-{3:3}, at: smc_switch_to_fallback+0x109/0x250 [smc]

 but task is already holding lock:
 ffff975246c8f918 (&ei->socket.wq.wait){..-.}-{3:3}, at: smc_switch_to_fallback+0xfe/0x250 [smc]

 other info that might help us debug this:
  Possible unsafe locking scenario:

        CPU0
        ----
   lock(&ei->socket.wq.wait);
   lock(&ei->socket.wq.wait);

  *** DEADLOCK ***

  May be due to missing lock nesting notation

 2 locks held by wrk/1391:
  #0: ffff975246040130 (sk_lock-AF_SMC){+.+.}-{0:0}, at: smc_connect+0x43/0x150 [smc]
  #1: ffff975246c8f918 (&ei->socket.wq.wait){..-.}-{3:3}, at: smc_switch_to_fallback+0xfe/0x250 [smc]

 stack backtrace:
 Call Trace:
  <TASK>
  dump_stack_lvl+0x56/0x7b
  __lock_acquire+0x951/0x11f0
  lock_acquire+0x27a/0x320
  ? smc_switch_to_fallback+0x109/0x250 [smc]
  ? smc_switch_to_fallback+0xfe/0x250 [smc]
  _raw_spin_lock_irq+0x3b/0x80
  ? smc_switch_to_fallback+0x109/0x250 [smc]
  smc_switch_to_fallback+0x109/0x250 [smc]
  smc_connect_fallback+0xe/0x30 [smc]
  __smc_connect+0xcf/0x1090 [smc]
  ? mark_held_locks+0x61/0x80
  ? __local_bh_enable_ip+0x77/0xe0
  ? lockdep_hardirqs_on+0xbf/0x130
  ? smc_connect+0x12a/0x150 [smc]
  smc_connect+0x12a/0x150 [smc]
  __sys_connect+0x8a/0xc0
  ? syscall_enter_from_user_mode+0x20/0x70
  __x64_sys_connect+0x16/0x20
  do_syscall_64+0x34/0x90
  entry_SYSCALL_64_after_hwframe+0x44/0xae

The nested locking in smc_switch_to_fallback() is considered to
possibly cause a deadlock because smc_wait->lock and clc_wait->lock
are the same type of lock. But actually it is safe so far since
there is no other place trying to obtain smc_wait->lock when
clc_wait->lock is held. So the patch replaces spin_lock() with
spin_lock_nested() to avoid false report by lockdep.

Link: https://lkml.org/lkml/2021/11/19/962
Fixes: 2153bd1 ("Transfer remaining wait queue entries during fallback")
Reported-by: syzbot+e979d3597f48262cb4ee@syzkaller.appspotmail.com
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Acked-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/smc/af_smc.c

@@ -585,7 +585,7 @@ static void smc_switch_to_fallback(struct smc_sock *smc, int reason_code)
 		 * to clcsocket->wq during the fallback.
 		 */
 		spin_lock_irqsave(&smc_wait->lock, flags);
-		spin_lock(&clc_wait->lock);
+		spin_lock_nested(&clc_wait->lock, SINGLE_DEPTH_NESTING);
 		list_splice_init(&smc_wait->head, &clc_wait->head);
 		spin_unlock(&clc_wait->lock);
 		spin_unlock_irqrestore(&smc_wait->lock, flags);